# rnn(initialHiddenStates:inputHiddenWeight:hiddenHiddenWeight:bias:direction:activation:outputSequence:) Adds an RNN operation to the current graph. ``` func rnn(initialHiddenStates: BNNSGraph.Builder.Tensor, inputHiddenWeight: BNNSGraph.Builder.Tensor, hiddenHiddenWeight: BNNSGraph.Builder.Tensor, bias: BNNSGraph.Builder.Tensor, direction: BNNSGraph.Builder.Direction, activation: BNNSGraph.Builder.Activation, outputSequence: Bool) -> (output: BNNSGraph.Builder.Tensor, hiddenStates: BNNSGraph.Builder.Tensor) ``` ## Parameters `initialHiddenStates` The initial hidden states, with the shape `(N, Hout)`, that the operation uses in the second matrix multiplication above when computing `h[0, ...]`. `inputHiddenWeight` The input-hidden weight with the shape `(Hout, Hin)`. `hiddenHiddenWeight` The hidden-hidden weight with the shape `(Hout, Hout)`. `bias` The bias (the sum of input-hidden and hidden-hidden biases) with the shape `(Hout,)`. `direction` An enumeration that specifies a forward or backward RNN. `activation` An enumeration that controls the output activation function. `outputSequence` When `true`, `output` is of shape `(L, N, Hout)` and contains hidden states from every step, `h[:, ...]`. When `false`, `output` is of shape `(1, N, Hout)` and contains hidden states from the last step, `h[-1, ...]`. ## Discussion For each time `t`, from `0` to `L-1`, this operation performs the following: ``` h[t, ...] = activation(matmul(x[t, ...], inputHiddenWeight^T) + matmul(h[t-1, ...], hiddenHiddenWeight^T) + bias) ``` The input tensor `x` is of shape `(L, N, Hin)`. `hiddenStates` is of shape `(N, Hout)` and contains hidden states from the last step, `h[-1, ...]`. --- Copyright © 2026 Apple Inc. All rights reserved. | [Terms of Use](https://www.apple.com/legal/internet-services/terms/site.html) | [Privacy Policy](https://www.apple.com/privacy/privacy-policy)